An LLM-Based Modeling and Decision Optimization for User-Centric Electric Vehicle Charging

Rapid adoption of Electric Vehicles (EVs) has significantly increased the demand for user-centric and personalized EV charging solutions. Existing charging decisions primarily rely on static optimization, involving complicated modeling and algorithms. This, however, poses a high communicative and technical barrier between users and machines, leading to a poor quality-of-experience. To this end, this paper proposes a novel large language model (LLM)-based multi-agent framework, acting as a mediator between users and devices to deliver a direct language-to-solution service. Our framework consists of three specialized agents, i.e., Agent-I, M, and -C, with each responsible for information querying, model construction, and code generation, separately. First, Agent-I extracts vehicle information and user preferences through interactive dialogues with users, and retrieves comprehensive brand and model parameters from external databases. Next, Agent-M filters out key modeling parameters, and grabs real-time electricity prices for charging. Finally, Agent-C generates the professional codes for commercial solvers, e.g., Gurobi, subsequently obtaining optimal charging solutions. Two cases with different EV brands and user preferences have then been tested to demonstrate that our framework can significantly improve user experience while maintaining the accuracy and optimality of charging decisions through a direct language-to-solution mode. This study showcases the potential of LLMs in smart charging systems and pave a way for user-centric resource optimization and scheduling.